Data Machina #217
Finetuning + RAG. Rotary Positional Embeddings. The Physics of LLMs. Wikipedia Search with Transformers. AgentSims. Llama 2 in WebLLM. GPT Pilot PoC. ChatDev. Knowledge Graph Prompting.
AI, FOMO and AIADH Disorder. There’s so much stuff happening in AI/ DL/ ML every week! There’s the overall feeling of FOMO, and that you must 100% learn every new AI stuff everyday. If you are not careful it’s easy to develop AIADHD (AI attention-deficit/hyperactivity disorder.) I guess the key is to stay focused, and remember that the fundamentals and first principles remain. Or not? Experiencing a bit of AIADHD can bring in serendipity and randomness too; which is why I’m sharing these 5 interesting AI snippets this week:
How to: Finetuning + RAG. Both finetuning and RAG are becoming super useful if you want to develop your own retrieval/ chat apps with an LLM finetuned with your own data. In this video you’ll learn how to best use finetuning with LLM apps, and combining it with RAG.
RoPE (Rotary positional embeddings) explained. This is an awesome video explainer. Unlike sinusoidal embeddings, RoPE are well behaved and more resilient to predictions exceeding the training sequence length. Modern LLMs have already steered away from sinusoidal embeddings for alternatives like RoPE. Learn about what's wrong with sinusoidal embeddings, the intuition of RoPE and how RoPE works.
Wolfram on the physics of LLMs. For the first time it seems there is a path to seamlessly converting natural language to computational language. Wolfram gives a great talk on the issues of physics and LLMs, and the way forward.
Wikipedia search-by-vibes. I love this. This is a browser-based search engine for Wikipedia. It uses sentence transformers to embed documents, product quantisation to compress embeddings, pq.js
to run distance computation in the browser, and transformers.js
to run sentence transformers in the browser for queries. Wikipedia search-by-vibes through millions of pages offline
AgentSims: A sandbox for LLM evaluation. AgentSims is an easy-to-use sandbox for researchers to evaluate LLMs. Researchers can build their evaluation tasks by adding agents and buildings on an interactive GUI or deploy and test new support mechanisms, with a few lines of code. Checkout the paper, demo, code, video: AgentSims: An Open-Source Sandbox for Large Language Model Evaluation.
Have a nice week.
10 Link-o-Troned
[free] Bootcamp: Data Structures & Optimisation for Fast Algos
Multi-Modal LLMs & Scaling Inference for 200TB Embeddings @Bytedance
the ML Pythonista
Deep & Other Learning Bits
AI/ DL ResearchDocs
data v-i-s-i-o-n-s
MLOps Untangled
AI startups -> radar
ML Datasets & Stuff
Postscript, etc
Tips? Suggestions? Feedback? email Carlos
Curated by @ds_ldn in the middle of the night.